Playback tutorial 8: Hardware-accelerated video decoding
Please port this tutorial to python!
Please port this tutorial to javascript!
Goal
Hardware-accelerated video decoding has rapidly become a necessity, as low-power devices grow more common. This tutorial (more of a lecture, actually) gives some background on hardware acceleration and explains how does GStreamer benefit from it.
Sneak peek: if properly setup, you do not need to do anything special to activate hardware acceleration; GStreamer automatically takes advantage of it.
Introduction
Video decoding can be an extremely CPU-intensive task, especially for higher resolutions like 1080p HDTV. Fortunately, modern graphics cards, equipped with programmable GPUs, are able to take care of this job, allowing the CPU to concentrate on other duties. Having dedicated hardware becomes essential for low-power CPUs which are simply incapable of decoding such media fast enough.
In the current state of things (June 2016) each GPU manufacturer offers a different method to access their hardware (a different API), and a strong industry standard has not emerged yet.
As of June 2016, there exist at least 8 different video decoding acceleration APIs:
-
VAAPI (Video Acceleration API): Initially designed by Intel in 2007, targeted at the X Window System on Unix-based operating systems, now open-source. It now also supports Wayland through dmabuf. It is currently not limited to Intel GPUs as other manufacturers are free to use this API, for example, Imagination Technologies or S3 Graphics. Accessible to GStreamer through the gstreamer-vaapi package.
-
OVD (Open Video Decode): Another API from AMD Graphics, designed to be a platform agnostic method for softrware developers to leverage the Universal Video Decode (UVD) hardware inside AMD Radeon graphics cards. Currently unavailable to GStreamer .
-
DCE (Distributed Codec Engine): An open source software library ("libdce") and API specification by Texas Instruments, targeted at Linux systems and ARM platforms. Accessible to GStreamer through the gstreamer-ducati plugin.
-
Android MediaCodec: This is Android's API to access the device's hardware decoder and encoder if available. This is accessible through the
androidmedia
plugin in gst-plugins-bad. This includes both encoding and decoding. -
Apple VideoTool Box Framework: Apple's API to access hardware is available through the
applemedia
plugin which includes both encoding through thevtenc
element and decoding through thevtdec
element. -
Video4Linux: Recent Linux kernels have a kernel API to expose hardware codecs in a standard way, this is now supported by the
v4l2
plugin ingst-plugins-good
. This can support both decoding and encoding depending on the platform.
Inner workings of hardware-accelerated video decoding plugins
These APIs generally offer a number of functionalities, like video decoding, post-processing, or presentation of the decoded frames. Correspondingly, plugins generally offer a different GStreamer element for each of these functions, so pipelines can be built to accommodate any need.
For example, the gstreamer-vaapi
plugin offers the vaapidecode
,
vaapipostproc
and vaapisink
elements that allow
hardware-accelerated decoding through VAAPI, upload of raw video frames
to GPU memory, download of GPU frames to system memory and presentation
of GPU frames, respectively.
It is important to distinguish between conventional GStreamer frames, which reside in system memory, and frames generated by hardware-accelerated APIs. The latter reside in GPU memory and cannot be touched by GStreamer. They can usually be downloaded to system memory and treated as conventional GStreamer frames when they are mapped, but it is far more efficient to leave them in the GPU and display them from there.
GStreamer needs to keep track of where these “hardware buffers” are though, so conventional buffers still travel from element to element. They look like regular buffers, but mapping their content is much slower as it has to be retrieved from the special memory used by hardware accelerated elements. This special memory types are negotiated using the allocation query mechanism.
This all means that, if a particular hardware acceleration API is
present in the system, and the corresponding GStreamer plugin is also
available, auto-plugging elements like playbin
are free to use
hardware acceleration to build their pipelines; the application does not
need to do anything special to enable it. Almost:
When playbin
has to choose among different equally valid elements,
like conventional software decoding (through vp8dec
, for example) or
hardware accelerated decoding (through vaapidecode
, for example), it
uses their rank to decide. The rank is a property of each element that
indicates its priority; playbin
will simply select the element that
is able to build a complete pipeline and has the highest rank.
So, whether playbin
will use hardware acceleration or not will depend
on the relative ranks of all elements capable of dealing with that media
type. Therefore, the easiest way to make sure hardware acceleration is
enabled or disabled is by changing the rank of the associated element,
as shown in this code:
static void enable_factory (const gchar *name, gboolean enable) {
GstRegistry *registry = NULL;
GstElementFactory *factory = NULL;
registry = gst_registry_get_default ();
if (!registry) return;
factory = gst_element_factory_find (name);
if (!factory) return;
if (enable) {
gst_plugin_feature_set_rank (GST_PLUGIN_FEATURE (factory), GST_RANK_PRIMARY + 1);
}
else {
gst_plugin_feature_set_rank (GST_PLUGIN_FEATURE (factory), GST_RANK_NONE);
}
gst_registry_add_feature (registry, GST_PLUGIN_FEATURE (factory));
return;
}
The first parameter passed to this method is the name of the element to
modify, for example, vaapidecode
or fluvadec
.
The key method is gst_plugin_feature_set_rank()
, which will set the
rank of the requested element factory to the desired level. For
convenience, ranks are divided in NONE, MARGINAL, SECONDARY and PRIMARY,
but any number will do. When enabling an element, we set it to
PRIMARY+1, so it has a higher rank than the rest of elements which
commonly have PRIMARY rank. Setting an element’s rank to NONE will make
the auto-plugging mechanism to never select it.
The GStreamer developers often rank hardware decoders lower than the software ones when they are defective. This should act as a warning.
Conclusion
This tutorial has shown a bit how GStreamer internally manages hardware accelerated video decoding. Particularly,
- Applications do not need to do anything special to enable hardware acceleration if a suitable API and the corresponding GStreamer plugin are available.
- Hardware acceleration can be enabled or disabled by changing the
rank of the decoding element with
gst_plugin_feature_set_rank()
.
It has been a pleasure having you here, and see you soon!
The results of the search are